Verified against GitHub Copilot code review · 2026-08-01
Steer Copilot's automated PR review at the risk that actually keeps you up at night
A PR-description addendum that redirects GitHub Copilot's automated code review toward the specific parts of a diff you are genuinely unsure about, instead of the roughly even pass it gives a diff by default.
The prompt
Ready to copy — highlighted parts are example details you can swap.
COPILOT REVIEW FOCUS WHAT THIS PR ACTUALLY DOES Adds a bulk-delete endpoint for admin users to remove multiple flagged comments at once. REVIEW THESE, IN THIS ORDER, WITH THE MOST SCRUTINY 1. Whether the bulk-delete endpoint checks that every comment ID belongs to the requesting admin's organization before deleting. 2. Whether the delete is wrapped in a transaction. 3. Rate limiting on the new endpoint. DECISIONS ALREADY MADE DELIBERATELY — DO NOT RE-RAISE THESE UNLESS YOU DISAGREE WITH THE REASONING We chose a hard delete over a soft-delete flag for this table, per the retention policy discussion in #482 — this was intentional. HOW DEEP TO GO ON THE TOP ITEM Go deep on item 1 specifically — trace the authorization check through every helper it calls, not just the top-level handler. Flag anything else in the diff too — an automated review that only ever looks where it was told to look eventually misses something a human reviewer would have caught by accident while reading in file order — but lead with the areas listed above and make sure every one of them gets an explicit answer, not just a comment if something happened to stand out. For each item in 1. Whether the bulk-delete endpoint checks that every comment ID belongs to the requesting admin's organization before deleting. 2. Whether the delete is wrapped in a transaction. 3. Rate limiting on the new endpoint., answer explicitly: does the diff actually handle this correctly, and if you are not certain, say so rather than defaulting to a comment only when something looks clearly wrong — an item you checked and found fine is worth a stated 'checked, looks correct,' because a reviewer reading this later needs to know the difference between reviewed and clean and not actually looked at closely. If a comment applies to a line only because of a broader pattern repeated across several places in this diff, say so once, name the pattern, and list every location it appears — do not leave a separate near-duplicate comment on each individual line the pattern shows up on; five copies of the same observation train a reader to skim past all five rather than absorb any of them. CONSTRAINTS - Do not restate We chose a hard delete over a soft-delete flag for this table, per the retention policy discussion in #482 — this was intentional. as a finding needing a decision — if you genuinely believe the reasoning behind one of them was wrong, say that explicitly and why, rather than raising it as a generic concern that reads as if the tradeoff had never been considered at all. - If 1. Whether the bulk-delete endpoint checks that every comment ID belongs to the requesting admin's organization before deleting. 2. Whether the delete is wrapped in a transaction. 3. Rate limiting on the new endpoint. asks about something the diff does not actually touch — for example a transaction-safety concern on a code path this PR does not modify — say so plainly rather than reviewing the unrelated existing code as if it were part of this change. - Rank your findings by actual severity, not by the order 1. Whether the bulk-delete endpoint checks that every comment ID belongs to the requesting admin's organization before deleting. 2. Whether the delete is wrapped in a transaction. 3. Rate limiting on the new endpoint. listed them in — the order in 1. Whether the bulk-delete endpoint checks that every comment ID belongs to the requesting admin's organization before deleting. 2. Whether the delete is wrapped in a transaction. 3. Rate limiting on the new endpoint. reflects what the author was worried about going in, which is not guaranteed to match what turned out to be the most serious issue once the diff was actually reviewed.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
GitHub's Copilot code review scans the full diff and leaves inline comments, but it has no independent way to know which part of a large diff the author is actually unsure about, so absent any signal it distributes attention roughly evenly across changed lines in something close to file order — which means the one function you actually lost sleep over gets the same amount of scrutiny as a one-line formatting change nearby, purely because they happen to sit in the same diff. Naming risk areas up front works on Copilot's review the same way it works on a human reviewer you hand a PR to and say 'please focus on X' — it does not add new information the model could not have derived on its own, but it changes where limited attention actually concentrates, and concentrated attention on the part that matters catches more than evenly distributed attention across the whole diff. Requiring an explicit answer for every named risk area, including a stated checked, looks fine rather than only a comment when something looks wrong, closes a specific gap in how automated review reads to a human afterward: silence on an item you asked about is genuinely ambiguous between reviewed and clean and the review never actually got to this, and only one of those should give you confidence to merge. Listing known tradeoffs already discussed suppresses a particular kind of noise that trains people to stop reading review comments at all — without it, an automated review will periodically re-raise a decision the team already made on purpose, indistinguishable in tone from a genuinely new finding, and a reviewer who has been burned by that a few times starts skimming past every comment on principle rather than reading each one for what it actually is. Consolidating a repeated pattern into one comment with every location listed, instead of one near-duplicate comment per occurrence, targets the same skim-training problem from a different angle: five copies of an identical observation are not five times more persuasive, they are the thing that teaches a reader Copilot's comments are noisy and safe to ignore in bulk.
Verified against
GitHub Copilot code review GA, 2026 (diff-scoped inline review) · 2026-08-01
Changelog
- 2026-08-01 — Initial publish, verified against GitHub Copilot code review GA.
Need this built into your business?
If a prompt isn't enough — custom software, built and maintained for you — that's Scult's day job.
EXPLORE CUSTOM SOFTWARE
